Welcome to Doom9's Forum, THE in-place to be for everyone interested in DVD conversion.

Before you start posting please read the forum rules. By posting to this forum you agree to abide by the rules.

 

Go Back   Doom9's Forum > Hardware & Software > Software players

Reply
 
Thread Tools Search this Thread Display Modes
Old 6th November 2015, 04:08   #34061  |  Link
x7007
Registered User
 
Join Date: Apr 2013
Posts: 315
Does anyone have issue with Potplayer and the new drivers nvidia 358.87 ? I have prender always at 0 ( GPU 95% with madvr and EVR , non CUVID or DXVA Back-to)

EDIT Never mind, I think something with Asus GPU Tweak 2 messed thing up, after restart everything ok including the overclock

Last edited by x7007; 6th November 2015 at 04:17. Reason: fixed
x7007 is offline   Reply With Quote
Old 6th November 2015, 04:15   #34062  |  Link
JarrettH
Registered User
 
Join Date: Aug 2004
Location: Canada
Posts: 860
I'd like to see more film examples too. Are there not any movie watchers here?

From Cube (1997)

Original at 100%
https://c2.staticflickr.com/6/5743/2...133b4b6a_o.png

xbr75 doubling, then Catmull + LL+AR downscaling
https://c1.staticflickr.com/1/746/22...6e5980ff_o.png

xbr100 doubling, then Catmull + LL+AR downscaling
https://c1.staticflickr.com/1/627/22...845b12c0_o.png

No upscaling refinement, sorry.

I think 100 is more faithful, but a tad too sharp.

Last edited by JarrettH; 6th November 2015 at 04:45.
JarrettH is offline   Reply With Quote
Old 6th November 2015, 06:58   #34063  |  Link
Warner306
Registered User
 
Join Date: Dec 2014
Posts: 1,127
Quote:
Originally Posted by JarrettH View Post
I'd like to see more film examples too. Are there not any movie watchers here?

From Cube (1997)

Original at 100%
https://c2.staticflickr.com/6/5743/2...133b4b6a_o.png

xbr75 doubling, then Catmull + LL+AR downscaling
https://c1.staticflickr.com/1/746/22...6e5980ff_o.png

xbr100 doubling, then Catmull + LL+AR downscaling
https://c1.staticflickr.com/1/627/22...845b12c0_o.png

No upscaling refinement, sorry.

I think 100 is more faithful, but a tad too sharp.
The ringing is much harder to detect in that image. If anything, the 100 seems to mask the ringing better than 75 due the extra sharpness.

But the image is definitely not offensive. It is very sharp, though, to be certain.
Warner306 is offline   Reply With Quote
Old 6th November 2015, 19:31   #34064  |  Link
seiyafan
Registered User
 
Join Date: Feb 2014
Posts: 162
For 1080 to 1440, has there any effort been done in comparing one step upscaling vs. doubling then downscaling? (Jinc vs. xbr/NNEDI3 then Catmull)
seiyafan is offline   Reply With Quote
Old 6th November 2015, 21:17   #34065  |  Link
har3inger
Registered User
 
Join Date: Feb 2014
Posts: 139
I did some comparisons with SuperRes and nnedi3 (32 neuron) to demonstrate the problem with Sharpness 1. All of these are also radius 0.66 and linear light unless otherwise indicated.

No SR (alignment off because nnedi3 introduces pixel shift) : http://i.imgur.com/wS8TZUw.png
Str 4 Sharp 1: http://i.imgur.com/q9Q0elg.png
Str 4 Sharp 2: http://i.imgur.com/Q9IP9JU.png
Str 4 Sharp 3: http://i.imgur.com/pQdX4EW.png
Str 4 Sharp 4: http://i.imgur.com/86MVCvM.png
Str 4 Sharp 1 Radius 0.33: http://i.imgur.com/zCoqUCg.png
Str 4 Sharp 1 Radius 0.33 (extremely obvious bugs on the old man): http://i.imgur.com/W2xMJJF.png

For all of these, the easiest area to spot the enhancements/bugs is around the white hat.

IMO, Sharp 2 is closest to the original in terms of edge enhancement. Sharp 1 looks broken. Sharp 3 and 4 are similar sharpness, and enhance the ringing beyond what's in the source.

Last edited by har3inger; 6th November 2015 at 21:23.
har3inger is offline   Reply With Quote
Old 6th November 2015, 23:58   #34066  |  Link
bozokaydin
Registered User
 
Join Date: Apr 2011
Posts: 16
Anyone can tell me what is the best setup for 1080p content? For 1080p content on 1080p screen does chroma superres work without upscaling? And i am a bit confused what is the difference in chroma superres and upscaling refinement superres?
__________________
HTPC:

Windows 8 64bit
Nvdia GTX 690
i7 3570k
Corsair H100 Water Cooling
Corsair 128 Gb Force SSD x 2 Raid 0+ Samsung 2 TBx 2 Hard Disc Raid 0
Creative Titanium HD
bozokaydin is offline   Reply With Quote
Old 7th November 2015, 01:05   #34067  |  Link
Warner306
Registered User
 
Join Date: Dec 2014
Posts: 1,127
Quote:
Originally Posted by seiyafan View Post
For 1080 to 1440, has there any effort been done in comparing one step upscaling vs. doubling then downscaling? (Jinc vs. xbr/NNEDI3 then Catmull)
There will be little to no difference unless you are using SuperRes. SuperRes is more effective when used with Image Doubling.
Warner306 is offline   Reply With Quote
Old 7th November 2015, 12:36   #34068  |  Link
aufkrawall
Registered User
 
Join Date: Dec 2011
Posts: 1,812
I've never seen a real difference between Jinc AR or image doubling for 1080p -> WQHD, neither with SuperRes. I'm watching every 1080p video upscaled to WQHD.

Jinc AR + SR (strength 3, sharpness 2, LL off):


superxbr 100 + SR:
aufkrawall is offline   Reply With Quote
Old 7th November 2015, 14:13   #34069  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by aufkrawall View Post
But are you sure that you put the right Adaptive Sharpen version into it?
It still gives me a very noticeable brightness increase with a strength of 0.5, which is not the case with the shaders linked by bacondither.
Are you sure you used the same strength in both? I've double checked and compared and if I setup the same strength, I get a similar increase in brightness in madVR's version and the latest experimental shaders. Please note that madVR runs the shaders in TV levels instead of PC levels, which means you may have to fine tune strength a bit to get the same effect. I think I used 1.0 strength in the shaders, and 0.9 in madVR, and got pretty much identical results.

The brightness increase totally vanishes if I run AdaptiveSharpen in linear light. So my impression is that the fix implemented by bacondither is not sufficient yet.

If you have a different opinion then please post comparison screenshots where using the shader HLSL files shows less brightness increase than madVR's implementation, although you've carefully matched sharpeness strength.

Quote:
Originally Posted by FreeFall View Post
Thanks madshi and nevcairiel, subtitle positioning is working well. The keep black bars visible if they contain subtitles option has stopped working in newer builds, none of the settings are working.
Are you sure? It seems to work fine here. If you can reproduce it, please send me a small sample of the video with which I can reproduce the problem on my PC, and tell me which exact checkboxes are set in the "zoom control" page.

Quote:
Originally Posted by Arm3nian View Post
I've watched LOTR 1000 times, but now I can watch it 1000 more times upside down! Sweet
Haha, yes! Will you lie upside down when watching it, though?

Quote:
Originally Posted by ryrynz View Post
Would be scaling, can't duplicate the issue with previous or latest build.
Ok, if you can't duplicate then let's ignore it for now.

Quote:
Originally Posted by KoD View Post
A feature request, if possible: some keyboard shortcut to switch the display refresh rate.
Maybe later. There's no easy way to implement it now.

Quote:
Originally Posted by Ver Greeneyes View Post
Is there a keyboard shortcut to toggle black bar detection or cropping? I don't need black bar detection that often, and there will probably always be videos on which it doesn't quite work, but I'd like to be able to enable it easily when the black bars are obvious and consistent.
Wouldn't it make more sense to try and perfect the detection so that it "always" works? Then you wouldn't ever have to switch it off. I don't really like if algos are unstable, so that you have to manually enable/disable them. I want madVR you to be able to configure madVR once and then never touch the settings again.

Quote:
Originally Posted by KoD View Post
I would welcome a proper implementation of frame duration display [more than ...]
Such an implementation already exists: Switch your display to 60Hz and enable smooth motion FRC.

Quote:
Originally Posted by SweetLow View Post
Umm... If this is my wish realization - it's work not as expected. In my case this variable always equal to the name of first device in devices list, not actual device of player window
Yes, that was meant to be your wish realization. Oh, will have to double check that.

Quote:
Originally Posted by pankov View Post
is it intentional that the hidden option to show the black bar detection using the ShowBlackBars empty file doesn't actually work if the "crop black bars" option is not enabled?
No, it's supposed to always work. Will check that.

Quote:
Originally Posted by pankov View Post
I'd like also to ask a question about the new rotation feature.
Is it intentional that the subtitles (both external and internal ones) are rotated together with the picture ... and moved almost to the center of the image?
Honestly, do you have any videos which need rotation *and* have subtitles?? I thought about subtitles, but it's pretty hard to implement correctly, and my thinking was that videos that need rotation are very unlikely to have subtitles. So I simply didn't spend any time on working on rotated subtitles. If you have legit files which need rotation and have subtitles, please let me know where they come from and why they are rotated. Then I'll think about whether it's worth spending time on...

Quote:
Originally Posted by AngelGraves13 View Post
Fair enough. I'd settle a simple deblocker that has 3 settings like the debanding.
Unfortunately there is no good deblocker available anywhere that I've seen. So I'd have to invest one myself. Maybe some day, but not soon.

Quote:
Originally Posted by kolak View Post
Unfortunately you are correct- many authoring studios don't use dithering at all have no clue what to do with it
Been in few big places and it was very disappointing.
Hope you're spreading the word where you can...

Quote:
Originally Posted by markanini View Post
Bug report: Once the interlaced flag is activated midstream MadVR sees it as interlaced for the remainder of the duration even if the stream switches back to reporting progressive.
Does EVR behave differently? Switching DXVA deinterlacing on/off is a complicated thing in madVR. Doing it all the time when the video switches those flags on/off could introduce a small stutter every time that happens. So I'm not sure if that's a good idea. How often does your stream switch back and forth?

Quote:
Originally Posted by mcn View Post
I'm attaching some screenshots regarding the "issue" with subtitles positioning.
The first screenshot is of the original video, 1916x1076, without black bars.
The other two screenshots are taken with the video shown in full screen at 1920x1200, where black bars are added.

Embedded subtitles are partially shown in the bottom black bar, while external ones aren't.
Both should be definitely rendered by XySubFilter since modifying the timing through its properties window affects both subs.
Hmmm... And both embedded and external ones are SRT? Can you maybe make a small sample of that video available to me, with the external subtitle file? (You can use PM for that.)

Quote:
Originally Posted by aufkrawall View Post
super-xbr has very big problems with ringing, you have to sacrifice a lot of sharpness to get to an acceptable ringing level if the source isn't optimal.
The last part is very important: "if the source isn't optimal". It's true that super-xbr "enhances" ringing artifacts that are already in the source, while NNEDI3 instead reduces them. That's one of the very nice things about NNEDI3. super-xbr should not introduce new ringing artifacts if the source is clean. But already existing ringing artifacts are enhanced the same way valid lines and details are enhanced. I'm not sure if it's fair to fault super-xbr for that.

Today many Blu-Rays are relatively clean with only low amount of ringing. I think for those super-xbr should work pretty well. If you are watching a lot of low quality content with strong ringing in it, then obviously NNEDI3 is the much better choice for that.

Quote:
Originally Posted by Warner306 View Post
I've read madshi prefers super-xbr to Jinc. And past posts have found NNEDI3 and super-xbr are the best algorithms to use with SuperRes because they are sharper.
Yes, that's true. I still like super-xbr quite a lot, especially for its quality-to-performance ratio. But aufkrawall has a valid point saying that super-xbr might not work so well for sources that have a lot of hard coded ringing in them. Maybe that will change if at some point I manage to find a good source deringing algorithm. Don't have one at this point, though.

Quote:
Originally Posted by aufkrawall View Post
sharpness values over 75 can introduce a lot of artifacts.
I wouldn't say that. Using sharpness values over 75 should not introduce artifacts. But it might enhance already existing artifacts. Which is a quite important difference.

Quote:
Originally Posted by Thunderbolt8 View Post
but if its really improved from a rather objective point of view then it would be a default setting, wouldnt it?
If every user had an NVidia Titan GPU then the default settings would be different to what they are now.

Quote:
Originally Posted by har3inger View Post
As far as I've seen, super-XBR doesn't really add much ringing from upscaling, but rather preserves source ringing a lot more than nnedi3 does. Also, in terms of sharpness, SuperXBR 75 is about the same as jinc3, so those two would be a fairer comparison for how badly each rings.

[...]

For clean sources, SuperXBR + SuperRes for 720p->1080p (CR downscale) has no ringing problems in my experience. I don't think it's fair to compare it to nnedi3, which is a lot better at cleaning up source ringing artifacts compared to any other madvr scaler.
Agreed.

Quote:
Originally Posted by aufkrawall View Post
I haven't seen a single case where SuperRes really introduces aliasing which isn't supposed to exist.
With this test image (which you found, IIRC), SuperRes does introduce a certain amount of aliasing:

http://madshi.net/SuperResTest.png

I've been using this test image to do some more SuperRes related tests. Here are my conclusions:

1) Linear light vs gamma light: If you downscale the image in LL, and then use SuperRes with GL, the lines get too thin. If you downscale the image in GL, and then use SuperRes with LL, the lines get too thick. The only way to reproduce the proper line thickness is to use either GL or LL for both downscaling and SuperRes. Which means that unfortunately I will probably not be able to remove the "linear light" SuperRes option, because the optimal value will not be the same for all sources.

2) Using linear light, different SuperRes sharpness values produce different artifacts. It seems that a sharpness value of 2 produces the least amount of artifacts.

3) A radius of 0.66 produces some aliasing. A radius of 1.00 means SuperRes adds no aliasing, but SuperRes also loses some of its effectiveness/sharpness. Values between 0.66 and 1.00 show some aliasing and some sharpness loss. I'm not sure which radius I should use. Maybe a compromise would be 0.80 or something like that?

Quote:
Originally Posted by har3inger View Post
A bigger problem seems to be the way small details with high contrast (pupils in eyes) get fairly messed up when sharpness is set lower than 2 for superres. In the blonde's eyeballs, the pupils get extra white dots in them that shouldn't be there. If it's anything like the images I've been testing, the problem fixes itself at sharpness 2 and 3 (but 3 tends to be pretty aliased).
Quote:
Originally Posted by har3inger View Post
I did some comparisons with SuperRes and nnedi3 (32 neuron) to demonstrate the problem with Sharpness 1. All of these are also radius 0.66 and linear light unless otherwise indicated.

No SR (alignment off because nnedi3 introduces pixel shift) : http://i.imgur.com/wS8TZUw.png
Str 4 Sharp 1: http://i.imgur.com/q9Q0elg.png
Str 4 Sharp 2: http://i.imgur.com/Q9IP9JU.png
Str 4 Sharp 3: http://i.imgur.com/pQdX4EW.png
Str 4 Sharp 4: http://i.imgur.com/86MVCvM.png
Str 4 Sharp 1 Radius 0.33: http://i.imgur.com/zCoqUCg.png
Str 4 Sharp 1 Radius 0.33 (extremely obvious bugs on the old man): http://i.imgur.com/W2xMJJF.png

For all of these, the easiest area to spot the enhancements/bugs is around the white hat.

IMO, Sharp 2 is closest to the original in terms of edge enhancement. Sharp 1 looks broken. Sharp 3 and 4 are similar sharpness, and enhance the ringing beyond what's in the source.
Thanks! These results match my own tests I've done today (see above for a summary of my test results).

I'm wondering whether I should remove the SuperRes sharpness setting again and hard code it to 2, since it seems to be the one which produces the most faithful reconstruction of the "ground truth"?
madshi is offline   Reply With Quote
Old 7th November 2015, 14:14   #34070  |  Link
Murmeltier
Registered User
 
Join Date: Oct 2015
Posts: 5
What are the advantages of "dithering in linear light"? Or more generally: What is linear light good for when it comes to dithering? What happens to the picture if I choose "don't dither in linear light" under "trade quality for performance"?
Murmeltier is offline   Reply With Quote
Old 7th November 2015, 14:22   #34071  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by Murmeltier View Post
What are the advantages of "dithering in linear light"? Or more generally: What is linear light good for when it comes to dithering? What happens to the picture if I choose "don't dither in linear light" under "trade quality for performance"?
In "devices -> your display -> properties" set your display to 3bit. Then compare dithering with linear light on/off, and you will see the difference.
madshi is offline   Reply With Quote
Old 7th November 2015, 14:25   #34072  |  Link
Ver Greeneyes
Registered User
 
Join Date: May 2012
Posts: 447
Quote:
Originally Posted by madshi View Post
Wouldn't it make more sense to try and perfect the detection so that it "always" works? Then you wouldn't ever have to switch it off. I don't really like if algos are unstable, so that you have to manually enable/disable them. I want madVR you to be able to configure madVR once and then never touch the settings again.
I agree, but I wonder if it's possible. I was watching a stream of Undertale the other day, a game that has a lot of black in it (the palette during battles is mostly black and white), and the algorithm seemed to get confused, occasionally cropping out even non-black areas. On another occasion, it cut off part of someone's webcam, which was partially overlaid on the game.

These videos aren't readily available since I had to download them off twitch and Youtube, but maybe I can make a reduced sample. Unfortunately that might involve large parts of the video to find where it goes wrong, which I'm not sure I have time for.

Last edited by Ver Greeneyes; 7th November 2015 at 14:27.
Ver Greeneyes is offline   Reply With Quote
Old 7th November 2015, 14:43   #34073  |  Link
madshi
Registered Developer
 
Join Date: Sep 2006
Posts: 9,140
Quote:
Originally Posted by SweetLow View Post
Umm... If this is my wish realization - it's work not as expected. In my case this variable always equal to the name of first device in devices list, not actual device of player window
Seems to work fine here. Are you sure?

Quote:
Originally Posted by Ver Greeneyes View Post
I agree, but I wonder if it's possible. I was watching a stream of Undertale the other day, a game that has a lot of black in it (the palette during battles is mostly black and white), and the algorithm seemed to get confused, occasionally cropping out even non-black areas. On another occasion, it cut off part of someone's webcam, which was partially overlaid on the game.

These videos aren't readily available since I had to download them off twitch and Youtube, but maybe I can make a reduced sample. Unfortunately that might involve large parts of the video to find where it goes wrong, which I'm not sure I have time for.
Hmmmm... How about setting up a profile then which enables black bar detection for videos which have a typical movie resolution (use "uncroppedSrcWidth" and "uncroppedSrcHeight"), e.g. for 1920x1080, 720x480 and 720x576 videos? That would cover PAL DVD, NTSC DVD and HD.
madshi is offline   Reply With Quote
Old 7th November 2015, 14:49   #34074  |  Link
leeperry
Kid for Today
 
Join Date: Aug 2004
Posts: 3,477
Quote:
Originally Posted by madshi View Post
I'm wondering whether I should remove the SuperRes sharpness setting again and hard code it to 2, since it seems to be the one which produces the most faithful reconstruction of the "ground truth"?
My real world tests also made me prefer St4/Sh2 from the get-go but in cases when sxbr25 is too soft and sxbr50 too sharp it seems somewhat reasonable to go St4/Sh1, of course I'd far prefer a sxbr strength slider.

It would also be great if chroma SR could get the same update as luma SR if any possible please.
leeperry is offline   Reply With Quote
Old 7th November 2015, 14:52   #34075  |  Link
onyx76
Registered User
 
onyx76's Avatar
 
Join Date: Aug 2015
Location: Argentina
Posts: 4
How to improve image quality?

Hello, anyone have any suggestions to improve my settings?

LAV filter decoder=software.
madVR DX11 FSE (1080P24/1080P60)
Output: 10 bit madVR + 12 bit Nvidia control panel -> Sony Bravia XBR-46HX929.

SD:
Chroma: Jinc + AR
Image: Jinc + AR
Double Luma: 1.2x or greater - NNEDI3 64 Neurons
Double Chroma: 1.2x or greater - NNEDI3 32 Neurons
Quadruple Luma: 2.4x or greater - NNEDI3 64 Neurons
Quadruple Chroma: 2.4x or greater - NNEDI3 32 Neurons
Downscaling: Catmull-Rom + AR + LL
Upscaling Refinement: Adaptive Sharpen (strength: 0.2)
Artifact Removal - Debanding: Low/High
Image Enhancements: Off
Dithering: ED2 + Use Colored Noise

720P:
Chroma: Jinc + AR
Image: Jinc + AR
Double Luma: 1.2x or greater - NNEDI3 64 Neurons
Double Chroma: Off
Downscaling: Catmull-Rom + AR + LL
Upscaling Refinement: Adaptive Sharpen (strength: 0.2)
Artifact Removal - Debanding: Low/High
Image Enhancements: Off
Dithering: ED2 + Use Colored Noise

1080P:
Chroma: Jinc + AR
Image: Jinc + AR
Double Luma: Off
Double Chroma: Off
Downscaling: Catmull-Rom + AR + LL
Upscaling Refinement: Adaptive Sharpen (strength: 0.2)
Artifact Removal - Debanding: Low/High
Image Enhancements: Off
Dithering: ED2 + Use Colored Noise


There is some advantage in using Super-xbr in Chroma?

What is better NNEDI3 + Adaptive Sharpen or NNEDI3 + SuperRes?
__________________
i5 OCed 4 Ghz, Zotac GTX 970 Amp! Extreme Oced (GPU 1382/1534 Mhz, Ram 8 Ghz), Windows 7 SP1 x64, madVR x64, MPC-HC x64, Sony Bravia 46".
onyx76 is offline   Reply With Quote
Old 7th November 2015, 14:52   #34076  |  Link
Ver Greeneyes
Registered User
 
Join Date: May 2012
Posts: 447
Quote:
Originally Posted by madshi View Post
Hmmmm... How about setting up a profile then which enables black bar detection for videos which have a typical movie resolution (use "uncroppedSrcWidth" and "uncroppedSrcHeight"), e.g. for 1920x1080, 720x480 and 720x576 videos? That would cover PAL DVD, NTSC DVD and HD.
That would work for DVDs, but I watch a lot of 1080p Youtube videos (1080p and 720p are both very common for Youtube and twitch). Anyway, I think it would be of more use to you if I just got you samples where it unexpectedly fails - I'll enable it and keep an eye out
Ver Greeneyes is offline   Reply With Quote
Old 7th November 2015, 14:53   #34077  |  Link
nlnl
Registered User
 
Join Date: Aug 2008
Posts: 176
madshi
My config: W7, AMD 15.2 driver
The problem: When source is video (29,97) and "force video mode' option is on, display does not switch to 59 rate, but swtches to 23.
If I set to 'video mode' using keyboard shortcut display switches.
nlnl is offline   Reply With Quote
Old 7th November 2015, 15:03   #34078  |  Link
markanini
Registered User
 
Join Date: Apr 2006
Posts: 299
Quote:
Originally Posted by madshi View Post
Does EVR behave differently? Switching DXVA deinterlacing on/off is a complicated thing in madVR. Doing it all the time when the video switches those flags on/off could introduce a small stutter every time that happens. So I'm not sure if that's a good idea. How often does your stream switch back and forth?
It was Intel QS decoding in LAV filters. Switching to DXVA2 or disabling hardware decoding solves the problem.
markanini is offline   Reply With Quote
Old 7th November 2015, 16:24   #34079  |  Link
agustin9
Registered User
 
Join Date: Aug 2008
Posts: 86
madshi, quick question, could Error Diffusion be doubling the cpu usage? Thanks!
agustin9 is offline   Reply With Quote
Old 7th November 2015, 18:03   #34080  |  Link
aufkrawall
Registered User
 
Join Date: Dec 2011
Posts: 1,812
Quote:
Originally Posted by madshi View Post
Are you sure you used the same strength in both? I've double checked and compared and if I setup the same strength, I get a similar increase in brightness in madVR's version and the latest experimental shaders. Please note that madVR runs the shaders in TV levels instead of PC levels, which means you may have to fine tune strength a bit to get the same effect. I think I used 1.0 strength in the shaders, and 0.9 in madVR, and got pretty much identical results.

The brightness increase totally vanishes if I run AdaptiveSharpen in linear light. So my impression is that the fix implemented by bacondither is not sufficient yet.

If you have a different opinion then please post comparison screenshots where using the shader HLSL files shows less brightness increase than madVR's implementation, although you've carefully matched sharpeness strength.
Gonna send you results via PM.

Quote:
Originally Posted by madshi View Post
The last part is very important: "if the source isn't optimal". It's true that super-xbr "enhances" ringing artifacts that are already in the source, while NNEDI3 instead reduces them. That's one of the very nice things about NNEDI3. super-xbr should not introduce new ringing artifacts if the source is clean. But already existing ringing artifacts are enhanced the same way valid lines and details are enhanced. I'm not sure if it's fair to fault super-xbr for that.

Today many Blu-Rays are relatively clean with only low amount of ringing. I think for those super-xbr should work pretty well. If you are watching a lot of low quality content with strong ringing in it, then obviously NNEDI3 is the much better choice for that.
Yeah, I said that super-xbr is mostly fine for BDs.
However, "mostly" is not "always" though. I don't know from what medium the Lighthouse example, which Warner306 posted, originates, but it shows some nasty haloing with super-xbr and it doesn't seem to be a crappy web video.

Quote:
Originally Posted by madshi View Post
I wouldn't say that. Using sharpness values over 75 should not introduce artifacts. But it might enhance already existing artifacts. Which is a quite important difference.
With a sharpness of 100, you get very fat lines compared to NNEDI3, which is less of a case with a sharpness of 75 (which is mostly softer than NNEDI3). Due to the line thickening, aliasing gets more visible with cartoons.

Quote:
Originally Posted by madshi View Post
With this test image (which you found, IIRC), SuperRes does introduce a certain amount of aliasing:

http://madshi.net/SuperResTest.png
This source however is already a bit aliased without any scaling (it has some ringing too).
I don't know about such fines structures. Maybe it is correct that they look aliased when they are upscaled, since this could be how the image would look like if it was created in that size in the beginning?
Maybe it's just a problem with very tiny pictures (or huge scaling factors).

Quote:
Originally Posted by madshi View Post
I've been using this test image to do some more SuperRes related tests. Here are my conclusions:

1) Linear light vs gamma light: If you downscale the image in LL, and then use SuperRes with GL, the lines get too thin. If you downscale the image in GL, and then use SuperRes with LL, the lines get too thick. The only way to reproduce the proper line thickness is to use either GL or LL for both downscaling and SuperRes. Which means that unfortunately I will probably not be able to remove the "linear light" SuperRes option, because the optimal value will not be the same for all sources.
Yeah, I did a ground truth comparison with a game screenshot with bright and dark fonts which lead me to the same conclusion.
However, with content "in the wild", I've found the thickening of LL SR more annoying than the brightness increase of gamma light in the meantime. Image gets brighter, but unlike traditional sharpen filters, it doesn't look "bleached" to me. And white fonts hardly become thicker with GL SR unless they are really on a very dark to black background, unlike black letters with LL SR.


Quote:
Originally Posted by madshi View Post
2) Using linear light, different SuperRes sharpness values produce different artifacts. It seems that a sharpness value of 2 produces the least amount of artifacts.
Yes. For this reason, I now use SR with the default values instead of strength 2 and sharpness 3. They end up mostly with the same sharpness and thanks to your optimizations, 3 passes are affordable now.

Quote:
Originally Posted by madshi View Post
3) A radius of 0.66 produces some aliasing. A radius of 1.00 means SuperRes adds no aliasing, but SuperRes also loses some of its effectiveness/sharpness. Values between 0.66 and 1.00 show some aliasing and some sharpness loss. I'm not sure which radius I should use. Maybe a compromise would be 0.80 or something like that?
0.8 looks already blurrier to me.
This can make sense however with very low-res content or high scaling factors. What about simple heuristics that decide the radius depending on the scaling factor? With a high scaling factor, the effect of SuperRes is much more perceivable, so the slight regression of sharpness wouldn't hurt.

Quote:
Originally Posted by madshi View Post
I'm wondering whether I should remove the SuperRes sharpness setting again and hard code it to 2, since it seems to be the one which produces the most faithful reconstruction of the "ground truth"?
Sharpness of 2 is definitely most useful.
Hm, but there could be the case when you want to use SuperRes, but can only afford 2 passes, and still you don't want to miss out on some additional sharpness. In this case, a sharpness of 3 could still make sense. Would probably still look better than additional LumaSharpen.
aufkrawall is offline   Reply With Quote
Reply

Tags
direct compute, dithering, error diffusion, madvr, ngu, nnedi3, quality, renderer, scaling, uhd upscaling, upsampling

Thread Tools Search this Thread
Search this Thread:

Advanced Search
Display Modes

Posting Rules
You may not post new threads
You may not post replies
You may not post attachments
You may not edit your posts

BB code is On
Smilies are On
[IMG] code is On
HTML code is Off

Forum Jump


All times are GMT +1. The time now is 14:04.


Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2024, vBulletin Solutions Inc.